Five-hour allowance drops before prompt submission, then drops further during a low-effort response
Environment
- ChatGPT Plus ($20/month)
- ChatGPT/Codex desktop on macOS
- Date: October 3, 2026; timezone: Europe/Istanbul
- Model reported by user: GPT-6.1 Sol, low/Light reasoning
- Separate Pi configuration supplied by user:
gpt-6.1-sol, reasoning low, service_tier = "default". This is Pi configuration, not a captured desktop request.
What happened
- I logged out of Pi.
- I opened the ChatGPT desktop application. My five-hour allowance showed 50% remaining.
- While typing my prompt, before pressing Enter, the allowance dropped to 38% remaining.
- After I submitted the prompt, the first usage query confirmed 62% used / 38% remaining.
- By the end of the response, the display showed 33% remaining.
These are two separate changes: 12 percentage points before submission, then 5 percentage points during the submitted request and response. The second change may include legitimate response usage; it does not explain the first change.
Evidence and limits
- A local token_count record at 01:56:39 reported 50% used / 50% remaining.
- A usage record at 02:02:45 reported 62% used / 38% remaining.
- The before-Enter timing and the final 33% reading are my direct observations; no screenshot was captured at those exact moments.
- The reviewed desktop startup log contained no turn/start before my first short question at 01:56. This is limited to the reviewed local log, not proof that no other account activity existed.
- A later 45-second interval without new model generation held steady at 65% used.
- Earlier Pi activity existed. Delayed accounting is possible, but local records cannot attribute the pre-submission debit to a particular request.
- Fast/priority usage is not established for the relevant requests. A local extension capable of overriding service_tier was found, but actual execution and transmitted request tier were not captured.
Expected behavior / requested investigation
Opening the app or typing an unsubmitted prompt should not itself initiate billable model work. If earlier completed work is being accounted for later, the usage display should make that clear.
Please correlate account-side request start/completion times, client origin, service tier, and allowance debits around 01:45–02:03 Europe/Istanbul. Check whether the pre-submission drop came from delayed accounting, background inference, another client, duplicate debits, or a display/accounting defect. Please provide attribution and correct any erroneous debits.
Account identifiers and raw logs are intentionally omitted from this public report and can be supplied privately through support.
Five-hour allowance drops before prompt submission, then drops further during a low-effort response
Environment
gpt-6.1-sol, reasoninglow,service_tier = "default". This is Pi configuration, not a captured desktop request.What happened
These are two separate changes: 12 percentage points before submission, then 5 percentage points during the submitted request and response. The second change may include legitimate response usage; it does not explain the first change.
Evidence and limits
Expected behavior / requested investigation
Opening the app or typing an unsubmitted prompt should not itself initiate billable model work. If earlier completed work is being accounted for later, the usage display should make that clear.
Please correlate account-side request start/completion times, client origin, service tier, and allowance debits around 01:45–02:03 Europe/Istanbul. Check whether the pre-submission drop came from delayed accounting, background inference, another client, duplicate debits, or a display/accounting defect. Please provide attribution and correct any erroneous debits.
Account identifiers and raw logs are intentionally omitted from this public report and can be supplied privately through support.